OpenCL 笔记5 reconstruction application

1. If the back-projection and the forward-projection operator are executed within separated kernels, therefore non-provided cache coherency of texture is not a problem.

2. The algorithm with a slower convergence can actually be computed much faster, with usage of the acceleration.

3. CUDA 1.1 the time consuming texture update operator dominates the overall reconstruction time. This can be dramatically relieved using 3-D texture arrays as introduced in CUDA 2.0.

4. ML-convex has a slower convergence than SART.